QNAP announces the TVS-AIh1688ATX, a business NAS with AI acceleration that features Intel Core Ultra processors with NPU and Intel Arc graphics, capable of reaching up to 36 TOPS for inference. Designed for image and video analytics, virtualization, multimedia post-production, and large-scale backups, this model combines high-performance computing with scalable storage and very high-speed networks, positioning itself as an edge platform for companies that want to bring AI and collaboration directly where data is generated.
QNAP’s offering isn’t just about adding CPU/GPU/NPU power: it includes 12 SATA bays for HDDs and 4 U.2 NVMe/SATA slots (PCIe Gen4 x2) for SSDs — useful as flash pools or cache —, along with 2 10GBASE-T ports and 2 2.5 GbE ports as standard, 3 PCIe Gen4 slots for upscaling to 25/100 GbE or added storage, USB 4, and Thunderbolt 5 support via an expansion card. Based on this hardware foundation, the QuTS hero (ZFS) operating system brings inline deduplication, WORM (immutability), ZIL power-loss protection, massive snapshots, and SnapSync for disaster recovery.
AI performance and integrated graphics: 36 TOPS for edge inference
The core of the TVS-AIh1688ATX can be configured with Intel Core Ultra 9 (24 cores / 24 threads, up to 5.6 GHz) or Core Ultra 7 (20 cores / 20 threads). Both feature NPU (Neural Processing Unit) and Intel Arc Graphics, adding AES-NI for accelerated encryption. QNAP caps the inference bandwidth at 36 TOPS combined, sufficient for vision models (detection, classification, segmentation) or transcoding acceleration directly within the NAS, eliminating the need for a dedicated GPU in many cases.
It’s key to understand the scope of that 36 TOPS: focused on inference (executing pre-trained models) and multimedia post-processing; training larger models still requires discrete accelerators. The advantage here is bringing AI close to where files and cameras reside, with minimal latency and controlled energy costs.
“Hybrid” storage architecture: 12+4 bays with enterprise-grade ZFS guarantees
The chassis provides 12 SATA bays for HDDs — offering elasticity in capacity and cost per TB — and 4 U.2 NVMe/SATA slots (PCIe Gen4 x2). This 12+4 setup enables creating high IOPS flash pools for virtual machines or metadata, while reserving HDDs or external JBODs for sequential and archival storage. QNAP promotes scale-out expansion to petabyte ranges via their catalog of JBOD shelves.
Regarding hardware, QuTS hero — QNAP’s ZFS-based variant — offers integrity and resilience such as:
- Auto-correction via checksums and copy-on-write.
- Inline deduplication and compression for space savings in environments with redundant data.
- WORM (Write Once, Read Many) for immutability — useful in compliance retention and ransomware protection.
- Almost unlimited snapshots and replication.
- Real-time SnapSync aiming for RPO = 0 between primary and secondary NAS when network latency allows.
Additional features include RAID-Z and advanced options like Triple Parity or Triple Mirror, meant to elevate fault tolerance over traditional RAID configurations. This model’s DDR5 ECC memory scales up to 192 GB to reinforce system consistency for 24/7 services.
Networking and expandability: standard 10 GbE, up to 100 GbE with PCIe
Out of the box, the TVS-AIh1688ATX features 2×10GBASE-T (10G/1G) and 2×2.5 GbE, supporting Port Trunking and failover. For higher loads or backup windows, it can be upgraded to 25 GbE or 100 GbE with PCIe Gen4 cards from QNAP or third parties, and additional M.2 via QM2 cards. For creative workflows and video streams, it includes 2×USB 4 Type-C (compatible with Thunderbolt 3/4) and supports Thunderbolt 5 with the QXP-T52P card — enabling up to four editing stations connected live to the NAS, with sustained I/O for 4K/8K and multi-user collaboration.
High availability and continuous operation: active-passive cluster
To minimize downtime, QNAP supports building HA clusters (active-passive) with two identical NAS units. If the primary node fails, the secondary seamlessly takes over. Combined with SnapSync, which performs block-level real-time replication, this strategy aims to reduce RTO and achieve RPO = 0. These capabilities are especially valued in virtualized environments, databases, or mission-critical applications.
Use cases: from edge AI to virtualization and large-scale backup
- Image/video analytics at the edge: object detection in IP cameras, counting, quality control on manufacturing lines, rapid tagging for data teams, with inferences powered by the NPU and Intel Arc.
- Studios and broadcasters: multi-camera ingest over 10 GbE/25 GbE, collaborative editing with Thunderbolt 5, assisted transcoding, and centralized storage under ZFS.
- Virtualization: multi-core CPUs, low-latency NVMe, and 10 GbE connectivity to host business service VMs and AI application containers (microservices, MLOps pipelines, inference APIs).
- Backup and archiving: deduplication, WORM, snapshots, and JBODs for long-term retention and disaster recovery. Immutability prevents malicious deletion or encryption of backups.
- Hybrid and cloud: myQNAPcloud One offers file and S3 object storage in one dashboard for external backups or capacity overflow.
Platform details and security
The combination of ECC DDR5 up to 192 GB and NVMe Gen4 x2 in U.2 minimizes metadata bottlenecks and VM latency. The AES-NI acceleration in Intel Core Ultra processors enables large-scale encryption without significant performance loss, relevant for TLS, encrypted volumes, or backup repositories. The enhancements in ZFS — such as checksums and self-healing — add a robust structural layer to prevent silent corruption and flash media failures.
What this means for the professional market
The TVS-AIh1688ATX embodies the hottest trend in “storage”: integrating AI compute with NAS units. For many organizations — government, industry, media, healthcare — moving petabytes to the cloud for inference is neither efficient nor immediate. Bringing AI closer to primary storage reduces latency and costs, eases compliance (data onsite), and enables real-time responses. The alternative — dedicated GPU servers — remains essential for training or large-scale inferencing, but a NAS with integrated NPU/GPU fills a much-demanded middle ground: pre-labeling, visual QA, semantic search in content, or supporting creative teams.
On the infrastructure side, the leap to 25/100 GbE with PCIe Gen4, the ZFS layer, and built-in HA elevate the TVS-AIh1688ATX to enterprise storage with mission-critical services, without sacrificing the ecosystem’s characteristic versatility. The inclusion of Thunderbolt 5 enables direct workflows from workstations with bandwidths once outside NAS scope.
Limitations and considerations
- The 36 TOPS do not replace a high-end discrete GPU for training or large-scale inference. They’re suited for edge tasks and preprocessing.
- For real RPO = 0 with SnapSync, latency between sites must be extremely low (metro environments) and bandwidth sufficient to cover the write delta.
- In multiuser 8K video environments, experience depends on network, NVMe pools, caches, and volume planning.
- Upscaling to 25/100 GbE and Thunderbolt 5 requires specific expansion cards and compatible networking hardware.
Quick specs
- CPU: up to Intel Core Ultra 9 (24C/24T, up to 5.6 GHz) or Core Ultra 7 (20C/20T).
- Accelerators: NPU integrated (up to 36 TOPS combined) + Intel Arc Graphics.
- Memory: DDR5 ECC up to 192 GB.
- Storage: 12×SATA (HDD) + 4×U.2 NVMe/SATA (PCIe Gen4 x2).
- Network: 2×10GBASE-T, 2×2.5 GbE; expansion to 25/100 GbE via PCIe Gen4.
- Expansion: 3×PCIe Gen4; USB 4; Thunderbolt 5 via QXP-T52P; JBOD for petabyte scale.
- OS: QuTS hero (ZFS) with deduplication, WORM, snapshots, and SnapSync.
Frequently Asked Questions
Who is the TVS-AIh1688ATX for?
It’s designed for businesses needing centralized storage with edge AI services, virtualization, and high-res video workflows. Particularly relevant in industry (quality control), retail (video analytics), media (collaborative 4K/8K editing), and IT (immutable repositories and RPO-sensitive disaster recovery).
What do the “36 TOPS” practically mean?
The TOPS (tera operations per second) of the NPU and integrated GPU focus on inference and on-site analytics and multimedia acceleration. They don’t replace discrete GPUs for training or large inference clusters, but they reduce latency and costs for edge tasks.
How is high availability and low RPO achieved?
Via an active-passive HA cluster with two NAS units and SnapSync in real time. When the primary fails, the secondary takes over (targeting RTO), and ongoing replication seeks RPO = 0 when network latency is very low and bandwidth is sufficient between sites.
Can I grow to 25/100 GbE and scale to petabytes?
Yes. The system supports PCIe Gen4 cards for 25/100 GbE NICs and can connect to QNAP JBODs for unlimited capacity growth, maintaining ZFS guarantees of integrity and resilience.
Sources: QNAP — Press release for the TVS-AIh1688ATX, product page, and hardware/software specs; documentation of QuTS hero (ZFS) and SnapSync.